DGT-TM: A freely Available Translation Memory in 22 Languages

Page created by Sandra Coleman
 
CONTINUE READING
DGT-TM: A freely Available Translation Memory in 22 Languages
    Ralf Steinberger♪, Andreas Eisele♫, Szymon Klocek♫, Spyridon Pilos♫ & Patrick Schlüter♫

                 (♪) European Commission                                            (♫) European Commission
               Joint Research Centre (JRC)                                   Directorate General for Translation (DGT)
                      Ispra (VA), Italy                                             Luxembourg, Luxembourg
             Ralf.Steinberger@jrc.ec.europa.eu                                  Firstname.Lastname@ec.europa.eu

                                                             Abstract
The European Commission’s (EC) Directorate General for Translation, together with the EC’s Joint Research Centre, is making
available a large translation memory (TM; i.e. sentences and their professionally produced translations) covering twenty-two official
European Union (EU) languages and their 231 language pairs. Such a resource is typically used by translation professionals in
combination with TM software to improve speed and consistency of their translations. However, this resource has also many uses for
translation studies and for language technology applications, including Statistical Machine Translation (SMT), terminology extraction,
Named Entity Recognition (NER), multilingual classification and clustering, and many more. In this reference paper for DGT-TM, we
introduce this new resource, provide statistics regarding its size, and explain how it was produced and how to use it.
Keywords: Translation Memory; 231 language pairs; European Commission.

                                                                     the EU’s internal market. Related developments are that the
          1. Introduction and Motivation                             European Institutions made publicly accessible the full-text
In recent years, the European Commission has released a              database of EU law EUR-Lex 2 , the database of EU
number of large-scale multilingual linguistic resources.             terminology IATE 3 , the multilingual wide-coverage
These are the parallel corpus JRC-Acquis (Steinberger et al.         thesaurus EuroVoc4, plus other resources for translators5
2006), the translation memory DGT-TM (2007), the named               (EC&DGT 2008). EUR-Lex provides free access to
entity recognition and normalisation resource JRC-Names              European Union law and other documents considered to be
(Steinberger et al. 2011) and the JRC Eurovoc Indexer JEX            public, written in all 23 official EU languages. The IATE
(Steinberger et al. 2012). The European Commission is                website (Inter-Active Terminology for Europe) gives access
now releasing an update to DGT-TM that contains                      to a database of EU inter-institutional terminology. IATE
documents published between the years 2004 and 2010 and              has been used in the EU institutions and agencies since
that is two times larger than the release from 2007. From            2004 for the collection, dissemination and shared
now on, it is planned to release DGT-TM data yearly. This            management of EU-specific terminology. EuroVoc is a
release, covering data produced until the end of the year            multilingual thesaurus originally built specifically for the
2010, is called DGT-TM Release 2011.                                 manual indexing and retrieval of multilingual documentary
    This effort to make multilingual data available and to           information of the EU institutions, but it is now much more
thus support the development of Language Technology                  widely used, e.g. by the libraries of many national
solutions is related to the affirmation that multilinguality is      governments in the EU. It is a multi-disciplinary thesaurus
one of the basic principles of the EU, guaranteeing cultural         covering fields that are sufficiently wide-ranging to
and linguistic diversity. By giving the EU citizen access to         encompass both Community and national points of view,
legislative and policy proposals in all the official EU              with a certain emphasis on parliamentary activities.
languages, translation and cross-lingual information access          EuroVoc is a controlled set of vocabulary which can also be
technology contribute to making the EU more transparent,             used outside the EU institutions, particularly by
egalitarian, accountable and democratic.                             parliaments. JRC has publicly released its JRC EuroVoc
    The release can furthermore be seen in the context of            indexer software JEX (Steinberger et al. 2012), which
Directive 2003/98/EC of the European Parliament and of               multi-label classifies documents according to the
the Council on the re-use of public sector information.1             multilingual EuroVoc thesaurus and thus allows
This directive recognises that public sector information             establishing links between documents written in different
such as multilingual collections of documents can be an              languages.
important primary material for digital content products and              While the original textual data contained in DGT-TM,
services, that their release may have an impact on                   produced by the European Institutions and the EU Member
cross-border exploitation of information, and that it may            States, consists of texts and their translation written for
                                                                     mostly legal purposes, the text collection has many
thus have a positive effect on an unhindered competition in
1                                                                    2
  For details and to read the full text of the regulation, see         See http://eur-lex.europa.eu/.
                                                                     3
http://eur-lex.europa.eu/LexUriServ/LexUriServ.do?uri=CELEX            See http://iate.europa.eu/.
                                                                     4
:32003L0098:EN:NOT. All URLs were last visited in October              See http://eurovoc.europa.eu/.
                                                                     5
2011.                                                                  See http://ec.europa.eu/dgs/translation/publications/.
practical uses outside the legal domain (see Section 2).             part-of-speech annotation, word sense disambiguation,
DGT’s Information Technology Unit and its Language                   and more (Yarowsky et al. 2001); Annotation
Applications Sector, who are responsible for the                     projection allows saving annotation time and it creates
development and maintenance of translation aids such as              more comparable resources for many languages;
machine translation, translation memories, and reference         • Cross-lingual plagiarism detection (Potthast et al.
and documentation search facilities (EC&DGT 2008),
                                                                     2011)
decided to make use of the EU’s text collections to enhance
their TMs. Now that the original full-text data has been         • Multilingual and cross-lingual clustering and
sentence-aligned and added to DGT’s TM, the resource is              classification (Wei et al. 2008).
being released publicly.                                         • More generally, creation of multilingual semantic
    We will now list possible Language Technology uses of            space in Lexical Semantic Analysis (LSA; Landauer &
multilingual TMs and similar resources (Section 2),                  Littman 1991), Kernel Canonical Correlation Analysis
describe DGT-TM and how it was produced (3), and give                (KCCA; Vinokourov et al. 2002), etc.
some practical details on its usage (4).                         When the full text (i.e. information on the ordering of the
                                                                 sentences in the document) is available, further uses are
            2. Possible uses of DGT-TM                           possible:
                                                                 • Translation studies, annotation projection for
The value of collections of parallel texts or sentences is
                                                                     co-reference      resolution,     discourse   analysis,
widely accepted and there have been efforts to make such
                                                                     comparative language studies;
collections available publicly (e.g. Macklovitch et al., 2000;
                                                                 • Checking translation consistency automatically;
Koehn, 2005), although some of them are only accessible
                                                                 • Making use of full-text information to improve SMT;
via web interfaces (e.g. Parasol, 2011). The number of
                                                                 • Testing and benchmarking alignment software (for
highly multilingual parallel collections is relatively low,
                                                                     sentences, words, etc.);
with EuroParl (Koehn 2005; 21 languages in the 2011
update), DGT-TM (2007) and JRC-Acquis (Steinberger et                        3. Details on DGT-TM-2011
al., 2006) (22 languages each) being among the most
highly multilingual ones. There have also been efforts to        In 2011, DGT decided to import a large number of official
make various third-party resources available via a single        EU documents with the purpose of adding them to the TM
website (Tiedemann, 2009).                                       used by DGT’s translators. For that purpose, the existing
    One area that crucially depends on parallel data is the      full-text document collections were automatically
creation of models for Statistical Machine Translation           sentence-aligned. The next sub-sections specify the text
(SMT). The initial work on SMT made use of proceedings           type of the documents, discuss the translation quality of the
of Canadian parliament debates (Hansard) available in            corpus, describe the alignment procedure and provide
English and French. Since 2001, SMT work funded by               statistics on the resulting multi-language-aligned corpus.
DARPA focused on translation from Chinese and Arabic
                                                                 3.1 Documents contained in DGT-TM-2011
into English, for which models were trained using large
parallel corpora, such as United Nations publications. The       Since January 1968 the EC’s Official Journal (OJ) has been
EuroParl corpus with its originally 11 languages allowed         published in two separate series: L (Legislation) and C
the creation of SMT systems for up to 110 language pairs         (Information and Notices). The documents selected for
(Koehn 2005) and provided a crucial pre-condition for            inclusion in DGT-TM-2011 are those that are part of the
work in projects like EuroMatrix and EuroMatrix Plus.6           L-Series, published between 2004 and 2010, inclusive. The
The publication of the JRC-Acquis in 2006 enabled the            documents are thus part of the Acquis Communautaire (i.e.
creation of SMT systems for 462 European language pairs          the common body of EU law). Documents that were
(Koehn et al. 2009).                                             already contained in the DGT-TM release of the year 2007
    The value of not only multi-monolingual, but parallel        were excluded, so as to avoid data overlap. New data
resources (corpora, dictionaries, tools) cannot be estimated     releases are planned annually. DGT also plans to process
high enough because it makes the effort to develop, train        the C-series documents in the future.
and test multilingual text mining tools more efficient and
comparable (Steinberger, 2011). Parallel collections of          3.2 Quality of the translations
sentences have been used, among other things, for the            DGT-TM contains official legal acts. Starting from the
following tasks:                                                 drafting language (in 2008, 72% of all documents were
• Producing multilingual lexical and semantic resources          drafted in English and 11% in French), the versions in the
     such as dictionaries and ontologies;                        other 22 languages are produced by highly qualified human
• Training and testing information extraction software;          translators specialised in specific subject domains. All
• Annotation projection across languages for Named               documents are linguistically and legally checked during a
     Entity Recognition (Ehrmann et al. 2011), sentiment         multi-step revision process. The quality is controlled at the
     analysis (Josef Steinberger et al., 2011),                  level of each translation service, by legal services and by
     multi-document summarisation (Turchi et al. 2010),          the EC’s Publications Office (PO). There is much focus on
                                                                 ensuring terminology consistency, which includes
6                                                                involvement of the public administrations of the EU
  See http://www.euromatrix.net/ and
                                                                 Member States. EU translators work in a cutting-edge IT
http://www.euromatrixplus.net/.
Table 1. Size of the data in DGT-TM Release 2011, covering the years 2004 until 2010. For comparison, the first column
 shows the number of TUs of the 2007 release. Size is expressed in number of TUs, words and characters (chars). The
 number of colons and semi-colons gives an approximate indication on how many TUs are sentence parts. The size on
 disk, expressed in Gibibytes, refers to the bilingual TMX files with English as the second language.

environment, with many custom-built enhancements aimed           be considered equivalent to sentences. We thus use both
at streamlining the work and ensuring quality and                terms synonymously.
consistency. While it is in the nature of translation that its       TUs in the parallel corpus were aligned between
quality is always arguable, it can be assumed that the           English and each of the other 21 languages. Alignments
translation quality in DGT-TM on average is of a very high       between language pairs not involving English are thus
standard.                                                        indirect and are produced by the accompanying software
                                                                 tool, exploiting the alignments of both languages with
3.3 Alignment of translation units                               English.
The manually translated full-text documents were thus the            DGT’s in-house alignment tool is tuned to dealing with
input to DGT’s processing, which resulted in a TM with a         the specific features of EC and EU documents. The
collection of sentences and their translations. Rather than      alignment algorithm makes use of anchors such as numbers
                                                                 and text numbering, as well as images and other
speaking of sentences as the basic alignment unit, it is more
                                                                 non-linguistic information. Text numbering is used to first
useful to speak of Translation Units (TUs). Translation          define zones, within which sentences can be aligned. The
Units are mostly traditional sentences, but they also contain    other clues are then used to strengthen the alignments. In
titles and section headings, and they may also be sentence       the absence of any external clues, the system uses character
parts separated by colons or semi-colons. The usage of           count statistics on the typical relative length of sentences
semi-colons to separate even longer, multi-paragraph             across languages.
sections is very common in legal jargon (between 16% and             Exact numbers for the alignment quality evaluation are
17% of all TUs; see also Table 1). However, for all              not available, but the alignment was tested in a production
intended usages of this resource, the TUs in DGT-TM can          setting where translators were confronted with the
automatically aligned translations. They were encouraged          even if they only want all the parallel sentences for a single
to notify any alignment errors, and where such errors were        language pair. Downloading only a subset of the zip files is
found, they were used to improve the alignment algorithm.         possible, but it will result in producing only a subset of the
Altogether, the alignment quality was found to be very            parallel corpus.
good, with only few errors.                                          The characters are encoded in UTF-16 Little Endian
                                                                  and the documents are in TMX format, which is a widely
3.4 Statistics about DGT-TM Version 2011                          used TM format. For backwards compatibility, the header
The result of the segmentation of all available documents         mentions TMX format 1.1, but the files are also compliant
in each language is a collection of about 38 million TUs in       with TMX 1.4b.
the 22 official EU languages (see Table 1). There are about
1.9 million TUs per language. Bulgaria and Romania                4.2 Extracting parallel sentences from
joined the EU only in 2007 and Maltese translations were          DGT-TM
not obligatory in the first three years of Malta’s EU
membership (2004-2007), which explains the smaller                In order to produce parallel sentence collections for any of
                                                                  the 231 possible language pairs involving the 22 input
number of TUs for these three languages. Irish Gaeilge
                                                                  languages, users need to use the included extraction
(GA) became the 23rd official EU language in 2007, but the        program TMXtract (with the extension .exe or .jar,
EU Institutions are currently exempt from the obligation to       depending on the operating system) and copy it into the
draft all acts in Irish.7 The number of Irish documents was       same directory as the zip files with the data. This software
too small to be included in this version of DGT-TM.               has not changed since DGT-TM (2007). The program is
    Table 2 shows the number of parallel TUs for each of          distributed in two versions: a version with graphical user
the 231 language pairs.                                           interface for the Windows operating system, consisting of
    Note that, while the DGT-TM releases of the years             two files: the program file and the library, and a
2007 and 2011 do not have any full documents in common,           machine-independent command line version in java byte
they do share many individual sentences. An analysis of           code that can be run on any machine supporting a Java
the English-Danish sets of sentences in both collections          runtime of version 1.4 or newer. Users can produce
                                                                  bilingual extractions for any language pair by following the
revealed that just over 3.5% of the combined sentences are
                                                                  explicit instructions available on the download site.
exact duplicates, which justifies the usage of TMs.
                                                                      Note that the sequence in the extracted files is not
However, TMs can even exploit partially matching
                                                                  necessarily the same as in the underlying documents, and
sentences so that translators can definitely benefit from
                                                                  redundancies of text segments like "Article 1" are
TMs of the size of DGT-TM. There are also duplicate TUs
                                                                  inevitable. The documents in the files are identified by the
within this new release. These are simply sentences or
                                                                  document number (Numdoc) of the original legislative
headers that occur repeatedly, across documents. While the
                                                                  document in the Eur-Lex database, but it should be noted
majority of these repeated TUs have been excluded when
                                                                  that these documents have been modified during the
creating this TM, others (especially from the earlier years)
                                                                  pre-processing steps.
are included: There are 2275 TUs that are identical across
all 22 languages, 13345 that are identical for ten or more
languages, and 330,158 TUs that are identical for any             4.3 Conditions of use
language pair.                                                    The DGT-TM database is the exclusive property of the
                                                                  European Commission. DGT-TM can be re-used
         4. Downloading and using DGT-TM                          and disseminated, free of charge, without the need to make
                                                                  an individual application, and both for commercial and
The resource and the software tool can be downloaded
                                                                  non-commercial purposes, within the limits set by the
from http://langtech.jrc.ec.europa.eu/JRC_Resources.html.         Commission Decision of 12 December 2011 on the re-use
The format of the DGT-TM-2011 files and the                       of Commission documents ("Re-use Decision"), published
accompanying software are identical to those of the               in the Official Journal of the European Union L330 of 14
DGT-TM release in 2007.                                           December 2011, pages 39 to 42.
                                                                       Any natural or legal person who re-uses the DGT-TM
4.1 Downloading DGT-TM                                            documents, in accordance with the conditions laid down in
The corpus has been split into 25 individual zip file             the Re-use Decision, is obliged to state the source of the
packages with up to 100MB each. Each zip file contains            documents used: the website address, the date of the latest
many TMX-files identified by the EUR-Lex identifier of            update and the fact that the European Commission retains
the underlying documents and a file list in plain text format     ownership of the data.
specifying the languages in which the documents are                    The software distributed with DGT-TM, necessary for
available. It is not necessary to unzip the files as the          its exploitation/extraction, must be used in accordance with
accompanying extraction program will access the data in           the conditions laid down in the licence GPL 2.0.
the zip files directly.                                                The database and the accompanying software are made
    The texts for the different languages are spread over the     available without any guarantee. The Commission is not
various zip files so that users will need to download all files   liable for any consequence stemming from the re-use.
                                                                  Moreover, the Commission is not liable for the quality of
7
    http://publications.europa.eu/code/pdf/370000en.htm#fn4-2
                                                                  the alignment nor the correctness of the data provided.
For more detailed information please refer to the              Parallel Corpora. Proceedings of the 8th International
section on the conditions of use of the DGT-TM website.            Conference Recent Advances in Natural Language
                                                                   Processing (RANLP'2011). Hissar, Bulgaria, 12-14
           5. Summary and future work                              September 2011.
Following the release of the JRC-Acquis (Steinberger et al.      Steinberger Ralf (2011). A survey of methods to ease the
2006) and the first version of the DGT-TM in 2007, DGT             development of highly multilingual Text Mining
and JRC have released, in 2012, a new and larger collection        applications. Language Resources and Evaluation
of sentences and their translations in up to 22 different          Journal. Online First on 11 October 2011.
languages. This latest release, named DGT-TM Version             Steinberger Ralf, Bruno Pouliquen, Anna Widiger,
2011, includes the documents from the L-Series                     Camelia Ignat, Tomaž Erjavec, Dan Tufiş, Dániel Varga
(Legislation) of the EU’s Official Journal published               (2006). The JRC-Acquis: A multilingual aligned parallel
between 2004 and 2010. It is planned to make future                corpus with 20+ languages. Proceedings of the 5th
releases of the TM annually.                                       International Conference on Language Resources and
    TMs consist of individual sentences and sentence-like          Evaluation (LREC'2006), pp. 2142-2147. Genoa, Italy,
fragments that do not allow reproducing the original text          24-26 May 2006.
while, for some purposes, it would be beneficial to have         Steinberger Ralf, Bruno Pouliquen, Mijail Kabadjov &
access to the full texts and thus to the order of sentences in     Erik van der Goot (2011). JRC-Names: A freely
the documents. It is therefore planned to also release a           available, highly multilingual named entity resource.
full-text version of the parallel documents in the same set        Proceedings of the 8th International Conference Recent
of languages.                                                      Advances      in    Natural     Language    Processing
                                                                   (RANLP'2011), pp. 104-110. Hissar, Bulgaria, 12-14
                      6. References                                September 2011.
DGT-TM (2007). http://langtech.jrc.ec.europa.eu/DGT-TM.html.     Steinberger Ralf, Mohamed Ebrahim & Marco Turchi
EC&DGT (2008). European Commission & Directorate                   (2012). JRC EuroVoc Indexer JEX - A freely available
  General for Translation – Translation Tools and                  multi-label categorisation tool. Proceedings of the 8th
  Workflow. Office for Official Publications, Brussels,            international conference on Language Resources and
  Belgium.                                                         Evaluation (LREC'2012), Istanbul, 21-27 May 2012.
Ehrmann Maud, Marco Turchi & Ralf Steinberger (2011).            Tiedemann Jörg (2009). News from OPUS - A Collection
  Building a Multilingual Named Entity-Annotated                   of Multilingual Parallel Corpora with Tools and
  Corpus. Proceedings of the 8th International Conference          Interfaces. In Recent Advances in Natural Language
  Recent Advances in Natural Language Processing                   Processing (vol V), pages 237-248, John Benjamins,
  (RANLP'2011). Hissar, Bulgaria, 12-14.9.2011.                    Amsterdam/Philadelphia.
Koehn Philipp (2005). EuroParl: A Parallel Corpus for            Turchi Marco, Josef Steinberger, Mijail Kabadjov & Ralf
  Statistical Machine Translation. Proceedings of Machine          Steinberger (2010). Using Parallel Corpora for
  Translation Summit.                                              Multilingual     (Multi-Document)       Summarisation
Koehn Philipp, Alexandra Birch & Ralf Steinberger (2009).          Evaluation. Multilingual and Multimodal Information
  462 Machine Translation Systems for Europe.                      Access Evaluation. Springer Lecture Notes for
  Proceedings of 12th Machine Translation Summit.                  Computer Science, LNCS 6360/2010, pp. 52-63.
Landauer Thomas & Michael Littman (1991). A Statistical          Vinokourov Alexei, John Shawe-Taylor, Nello Cristianini
  Method for Language-Independent Representation of                (2002). Inferring a semantic representation of text via
  the Topical Content of Text Segments. Proceedings of             cross-language correlation analysis. Advances of Neural
  the 11th International Conference ‘Expert Systems and            Information Processing Systems 15.
  Their Applications’, vol. 8: pp. 77-85.                        Wei Chih-Ping, Christopher C. Yang, Chia-Min Lin (2008).
Macklovitch Elliott, Michel Simard & Philippe Langlais             A Latent Semantic Indexing-based approach to
  (2000). TransSearch: A Free Translation Memory on the            multilingual document clustering. Decision Support
  World Wide Web. Proceedings of LREC.                             Systems 45 (2008), 606–620.
Parasol (2011). Parasol – A parallel corpus of Slavic and        Yarowsky, D., Ngai, G. and Wicentowski, R. (2001)
  other languages. http://parasol.unibe.ch/.                       Inducing Multilingual Text Analysis Tools via Robust
Potthast Martin, Alberto Barrón-Cedeño, Benno Stein,               Projection across Aligned Corpora. In Proceedings of
  Paolo Rosso (2011). Cross-Language Plagiarism                    HLT’01, San Diego.
  Detection. In: Language Resources and Evaluation.
  Special Issue on Plagiarism and Authorship Analysis,
  vol. 45, num. 1, pp.45-62.
Steinberger Josef, Polina Lenkova, Mijail Kabadjov, Ralf
  Steinberger & Erik van der Goot (2011). Multilingual
  Entity-Centered Sentiment Analysis Evaluated by
Table 2. Number of TUs per language pair.
You can also read